feat: add GLM-5 SFT loss mask support

[stevewx]

Summary

• Add --loss-mask-type glm5 for multi-turn SFT on GLM-5 models
• GLM-5 uses role-token delimiters (<|assistant|>, <|user|>, etc.) with no closing tags, unlike Qwen's ChatML <|im_start|>/<|im_end|> format
• Handles both </think> (non-thinking) and <think>...</think> (thinking) modes
• Supports step_loss_mask per-turn control and tool-calling via <|observation|> role

Changes

• slime/utils/mask_utils.py — add gen_multi_turn_loss_mask_glm5() method using char-level offset mapping
• slime/utils/arguments.py — add "glm5" to --loss-mask-type choices
• tests/utils/test_loss_mask_type_glm5.py — 7 unit tests covering single-turn, multi-turn, step_loss_mask, tool calling, tool schemas, no-system-message, and length consistency

Test plan

• 7 unit tests pass locally (pytest tests/utils/test_loss_mask_type_glm5.py)
• Existing Qwen3.5 tests unaffected (pytest tests/utils/test_loss_mask_type_qwen35.py)
• Verified against real GLM-5 tokenizer (zai-org/GLM-5)
• Tested GLM-5 SFT training with OpenHermes-2.5